We consider neural networks of both types, unsupervised and supervised learning models. In the first case we review extensions of the Self-Organizing Map (SOM) and the Neural Gas (NG) which can be used for learning of relevant informations contained in the data (structure adaptaion, magnification control). This can be taken as a kind of knowledge discovery in data mining. In the latter case of supervised learning we will consider the task to discover what data/structures are relevant to obtain a given classfication. In particular, we focus on determining relevant input dimensions in data according to a given classification what we call relevance learning. We give views to some recent developments in generalization of well known vector quantization algorithms as LVQ and NG.